Our site shows a "strongest combinations" panel: the two, three and four highest-rated selections of the day. It is the most-clicked thing we publish and the most misunderstood, so here is the arithmetic in full, including the part that makes it look bad.
Probabilities of independent events multiply. Two 70% picks together are 0.70 × 0.70 = 49% — already a coin flip. Three are 34%. Four are 24%. Nothing has gone wrong at that point: four excellent selections have combined into a bet that loses three times out of four, purely by arithmetic. The picks did not get worse; there are simply more ways to fail.
This is why our combinations panel prints the multiplied figure in bold next to every set. A punter looking at four confident green rows has an intuition of "very likely". The intuition is wrong by a factor of three, and no amount of pick quality repairs it.
We freeze the day's combinations into a version-controlled ledger each morning, before kick-off, and grade them publicly. Across the first fortnight the strongest pairs landed 5 of 12 for overs and 5 of 12 for unders — close to the ~50% the maths predicts. The strongest fours landed 2 of 9 and 0 of 11. Both figures are consistent with expectation. Neither is a good bet.
There is a further wrinkle we state plainly: multiplication assumes the games are independent. Football selections often are not. If our model is systematically running hot on a given weekend — a hot league anchor, an unusual weather pattern, a fixture list full of derbies — the errors correlate, and the true combined probability is lower than the multiplied figure. Correlated errors make accumulators worse than the arithmetic suggests, never better.
We publish the panel because people want it and because the honest version is more useful than the marketing version. What we will not do is describe four multiplied selections as "strong" without printing the 24% beside it.
Method: walk-forward testing — every prediction is made using only information available before kick-off, then graded against what happened. Studies run June-July 2026 on the site's own dataset. Questions: contact us.